ISO 42001 AI Governance Audit

ISO 42001 AI治理审计

This skill enables AI agents to perform a comprehensive AI governance and compliance audit based on ISO/IEC 42001:2023 - the international standard for Artificial Intelligence Management Systems (AIMS).

ISO 42001 provides a framework for responsible development, deployment, and use of AI systems, addressing risks, ethics, security, transparency, and regulatory compliance.

Use this skill to ensure AI projects follow international best practices, manage risks effectively, and maintain ethical standards throughout the AI lifecycle.

Combine with security audits, code reviews, or ethical AI assessments for comprehensive AI system evaluation.

本技能支持AI Agent基于ISO/IEC 42001:2023——人工智能管理体系（AIMS）的国际标准，开展全面的AI治理与合规审计。

ISO 42001为AI系统的负责任开发、部署与使用提供框架，涵盖风险、伦理、安全、透明度及监管合规等维度。

使用本技能可确保AI项目遵循国际最佳实践，有效管理风险，并在AI全生命周期内维持伦理标准。

可结合安全审计、代码审查或伦理AI评估，实现AI系统的全面评估。

When to Use This Skill

何时使用本技能

Invoke this skill when:

Developing or integrating AI systems
Ensuring AI governance and compliance
Managing AI risks and ethical concerns
Preparing for AI regulatory requirements (EU AI Act, etc.)
Auditing existing AI implementations
Establishing AI governance frameworks
Responding to AI security or bias incidents
Planning responsible AI deployment
Documenting AI systems for stakeholders

在以下场景调用本技能：

开发或集成AI系统时
需确保AI治理与合规性时
管理AI风险与伦理问题时
为AI监管要求（如《欧盟AI法案》等）做准备时
审计现有AI实现时
建立AI治理框架时
应对AI安全或偏见事件时
规划负责任的AI部署时
为利益相关者记录AI系统时

Inputs Required

所需输入

When executing this audit, gather:

ai_system_description: Detailed description (purpose, capabilities, data used, users affected, deployment context) [REQUIRED]
use_case: Specific application (e.g., hiring tool, medical diagnosis, content moderation) [REQUIRED]
risk_category: High-risk, limited-risk, or minimal-risk per EU AI Act classification [OPTIONAL but recommended]
existing_documentation: Technical docs, data sheets, model cards, risk assessments [OPTIONAL]
stakeholders: Who develops, deploys, uses, and is affected by the AI [OPTIONAL]
regulatory_context: Applicable laws (GDPR, EU AI Act, industry regulations) [OPTIONAL]

执行审计时，需收集以下信息：

ai_system_description：AI系统的详细描述（用途、功能、所用数据、受影响用户、部署环境）【必填】
use_case：具体应用场景（如招聘工具、医疗诊断、内容审核）【必填】
risk_category：依据《欧盟AI法案》分类的高风险、有限风险或低风险【可选但推荐】
existing_documentation：技术文档、数据表、模型卡片、风险评估报告【可选】
stakeholders：AI系统的开发者、部署者、使用者及受影响方【可选】
regulatory_context：适用法规（GDPR、《欧盟AI法案》、行业特定法规）【可选】

ISO 42001 Framework Overview

ISO 42001框架概述

ISO 42001 is structured around 10 key clauses plus supporting annexes:

ISO 42001由10个核心条款及配套附录构成：

Core Clauses

核心条款

Scope - Define AIMS boundaries
Normative References - Related standards
Terms and Definitions - AI terminology
Context of Organization - Internal/external factors
Leadership - Management commitment and roles
Planning - Objectives and risk management
Support - Resources, competence, communication
Operation - AI system lifecycle management
Performance Evaluation - Monitoring and measurement
Improvement - Continual enhancement

范围 - 定义AIMS的边界
规范性引用文件 - 相关标准
术语与定义 - AI相关术语
组织环境 - 内外部影响因素
领导力 - 管理层承诺与角色
规划 - 目标设定与风险管理
支持 - 资源、能力与沟通
运行 - AI系统全生命周期管理
绩效评价 - 监控与测量
改进 - 持续优化

Key ISO 42001 Principles

ISO 42001核心原则

1. Risk-Based Approach

1. 基于风险的方法

Identify, assess, and mitigate AI-specific risks
Consider technical, ethical, legal, and social risks
Proportionate controls based on risk level

识别、评估并缓解AI特定风险
考量技术、伦理、法律与社会风险
根据风险等级匹配相应控制措施

2. Ethical AI

2. 伦理AI

Fairness and non-discrimination
Transparency and explainability
Human oversight and control
Privacy and data protection
Accountability

公平与非歧视
透明度与可解释性
人类监督与控制
隐私与数据保护
问责制

3. Lifecycle Management

3. 全生命周期管理

Design → Development → Deployment → Monitoring → Decommissioning
Continuous evaluation and improvement
Documentation throughout

设计 → 开发 → 部署 → 监控 → 退役
持续评估与改进
全流程文档记录

4. Stakeholder Engagement

4. 利益相关者参与

Involve affected parties
Clear communication about AI use
Mechanisms for feedback and redress

纳入受影响方
清晰沟通AI使用情况
建立反馈与申诉机制

Audit Procedure

审计流程

Follow these steps systematically:

系统遵循以下步骤开展审计：

Step 1: Context and Scope Analysis (15 minutes)

步骤1：环境与范围分析（15分钟）

Understand the AI System:

Define AIMS Scope (Clause 4)
- What AI systems are included?
- Organizational boundaries
- Interfaces with other systems
- Exclusions (if any)
Identify Stakeholders:
- Developers: Who builds the AI?
- Deployers: Who operates it?
- Users: Who interacts with it?
- Affected Parties: Who is impacted by decisions?
- Regulators: What oversight exists?
Assess Context:
- Industry and domain
- Regulatory environment (EU AI Act, GDPR, sector-specific)
- Cultural and social considerations
- Technical maturity and capabilities
Risk Classification (EU AI Act alignment):
- Unacceptable Risk: Prohibited uses (e.g., social scoring, real-time biometric surveillance)
- High Risk: Significant impact (e.g., employment, credit scoring, healthcare, law enforcement)
- Limited Risk: Transparency obligations (e.g., chatbots, deepfakes)
- Minimal Risk: Low impact (e.g., spam filters, recommender systems)

了解AI系统：

定义AIMS范围（条款4）
- 涵盖哪些AI系统？
- 组织边界
- 与其他系统的接口
- 排除项（如有）
识别利益相关者：
- 开发者：谁构建AI系统？
- 部署者：谁运营AI系统？
- 使用者：谁与AI系统交互？
- 受影响方：谁会受AI决策影响？
- 监管机构：存在哪些监管机制？
评估环境：
- 行业与领域
- 监管环境（《欧盟AI法案》、GDPR、行业特定法规）
- 文化与社会因素
- 技术成熟度与能力
风险分类（对齐《欧盟AI法案》）：
- 不可接受风险：被禁止的用途（如社会评分、实时生物特征监控）
- 高风险：影响重大（如就业、信用评分、医疗、执法）
- 有限风险：需履行透明度义务（如聊天机器人、深度伪造）
- 低风险：影响较小（如垃圾邮件过滤器、推荐系统）

Step 2: Leadership and Governance Evaluation (20 minutes)

步骤2：领导力与治理评估（20分钟）

Clause 5: Leadership

条款5：领导力

Step 3: Planning and Risk Management (30 minutes)

步骤3：规划与风险管理（30分钟）

Clause 6: Planning

条款6：规划

6.1 Actions to Address Risks and Opportunities

ISO 42001 Risk Categories:

Technical Risks
- Model accuracy and reliability
- Robustness to adversarial attacks
- Data quality and bias
- System failures and errors
- Integration issues
- Scalability and performance
Ethical Risks
- Discrimination and bias
- Lack of fairness
- Privacy violations
- Lack of transparency
- Autonomy and human dignity impacts
Legal and Compliance Risks
- Regulatory non-compliance (GDPR, EU AI Act)
- Intellectual property issues
- Liability for AI decisions
- Contractual obligations
Operational Risks
- Dependency on AI vendors
- Skills and competency gaps
- Change management failures
- Inadequate monitoring
Reputational Risks
- Public trust erosion
- Media scrutiny
- Stakeholder backlash
- Brand damage from AI failures

Risk Assessment Process:

For each identified risk:

markdown

undefined

6.1 应对风险与机遇的行动

ISO 42001风险类别：

技术风险
- 模型准确性与可靠性
- 对抗攻击鲁棒性
- 数据质量与偏见
- 系统故障与错误
- 集成问题
- 可扩展性与性能
伦理风险
- 歧视与偏见
- 缺乏公平性
- 隐私侵犯
- 缺乏透明度
- 自主性对人类尊严的影响
法律与合规风险
- 违反监管要求（GDPR、《欧盟AI法案》）
- 知识产权问题
- AI决策的责任归属
- 合同义务
运营风险
- 依赖AI供应商
- 技能与能力差距
- 变更管理失败
- 监控不足
声誉风险
- 公众信任受损
- 媒体审视
- 利益相关者反对
- AI故障导致品牌受损

风险评估流程：

针对每个识别出的风险：

markdown

undefined

Risk: [Name]

风险：[名称]

Category: Technical / Ethical / Legal / Operational / Reputational Likelihood: Low / Medium / High Impact: Low / Medium / High / Critical Risk Level: [Likelihood × Impact]

Description: [What could go wrong] Affected Stakeholders: [Who is impacted] Existing Controls: [Current mitigations] Residual Risk: [Risk after controls]

Treatment Plan:

Accept (if low risk)
Mitigate (reduce likelihood/impact)
Transfer (insurance, contracts)
Avoid (don't deploy feature)

Mitigation Actions:

[Specific action 1]
[Specific action 2]
[Specific action 3]

Owner: [Who is responsible] Timeline: [When to implement] Review Date: [When to reassess]


**Example Risks:**

**Risk 1: Algorithmic Bias in Hiring AI**
- Category: Ethical, Legal
- Likelihood: High (historical bias in training data)
- Impact: Critical (discrimination, legal liability)
- Risk Level: **CRITICAL**
- Mitigation:
  - Bias testing on protected attributes
  - Diverse training data
  - Regular fairness audits
  - Human review of decisions
  - Transparent criteria documentation

**Risk 2: Data Poisoning Attack**
- Category: Technical, Security
- Likelihood: Medium (if public data sources)
- Impact: High (model corruption)
- Risk Level: **HIGH**
- Mitigation:
  - Data validation and sanitization
  - Anomaly detection
  - Provenance tracking
  - Regular model retraining
  - Adversarial testing

---

**6.2 AI Objectives and Planning to Achieve Them**

**Evaluate:**
- [ ] Measurable AI objectives defined
- [ ] Aligned with organizational goals
- [ ] Consider stakeholder needs
- [ ] Include ethical and safety criteria
- [ ] Resources and timelines allocated
- [ ] Performance indicators established

**SMART AI Objectives Example:**
- "Achieve 95% accuracy while maintaining <5% false positive rate across all demographic groups by Q4"
- "Reduce bias disparity in loan approvals to <2% between groups by 2026"
- "Maintain 100% compliance with GDPR data subject rights"

---

类别：技术 / 伦理 / 法律 / 运营 / 声誉 可能性：低 / 中 / 高影响：低 / 中 / 高 / 严重 风险等级：[可能性 × 影响]

描述：[可能出现的问题] 受影响利益相关者：[受影响方] 现有控制措施：[当前缓解手段] 剩余风险：[采取控制措施后的风险]

处理方案：

接受（若为低风险）
缓解（降低可能性/影响）
转移（保险、合同）
避免（不部署该功能）

缓解行动：

[具体行动1]
[具体行动2]
[具体行动3]

负责人：[责任人] 时间线：[实施时间] 复查日期：[重新评估时间]


**风险示例：**

**风险1：招聘AI中的算法偏见**
- 类别：伦理、法律
- 可能性：高（训练数据存在历史偏见）
- 影响：严重（歧视、法律责任）
- 风险等级：**严重**
- 缓解措施：
  - 针对受保护属性进行偏见测试
  - 使用多样化训练数据
  - 定期开展公平性审计
  - 人工审查决策结果
  - 透明记录评估标准

**风险2：数据投毒攻击**
- 类别：技术、安全
- 可能性：中（若使用公共数据源）
- 影响：高（模型损坏）
- 风险等级：**高**
- 缓解措施：
  - 数据验证与清理
  - 异常检测
  - 来源追踪
  - 定期重新训练模型
  - 对抗性测试

---

**6.2 AI目标与实现规划**

**评估：**
- [ ] 已定义可衡量的AI目标
- [ ] 与组织目标对齐
- [ ] 考虑利益相关者需求
- [ ] 包含伦理与安全标准
- [ ] 已分配资源与时间线
- [ ] 建立绩效指标

**SMART AI目标示例：**
- "到第四季度实现95%的准确率，同时所有群体的假阳性率<5%"
- "到2026年将贷款审批中的偏见差异降至<2%"
- "100%符合GDPR数据主体权利要求"

---

Step 4: Support and Resources (20 minutes)

步骤4：支持与资源（20分钟）

Clause 7: Support

条款7：支持

Step 5: Operation - AI Lifecycle Management (40 minutes)

步骤5：运行 - AI全生命周期管理（40分钟）

Clause 8: Operation

条款8：运行

8.1 Operational Planning and Control

ISO 42001 requires managing AI through its entire lifecycle:

AI Lifecycle Stages:

Design → Development → Validation → Deployment → Monitoring → Maintenance → Decommissioning

STAGE 1: Design and Requirements

Evaluate:

Clear problem definition and success criteria
Stakeholder needs assessed
Ethical considerations identified early
Regulatory requirements mapped
Feasibility and impact analysis conducted
Alternatives to AI considered

Questions:

Is AI the right solution, or could simpler approaches work?
What could go wrong?
Who is affected and how?
What data is needed and available?
What are the ethical red lines?

Red Flags:

Using AI for high-stakes decisions without justification
No clear success metrics
Ignoring stakeholder concerns
Insufficient data or biased data sources

STAGE 2: Data Management

Evaluate:

Data quality assessed (accuracy, completeness, timeliness)
Bias and representativeness analyzed
Data sources documented and verified
Privacy and consent requirements met
Data security and access controls
Data minimization principles applied

Data Quality Dimensions:

Accuracy: Correct and error-free
Completeness: No missing values in critical fields
Consistency: Uniform across sources
Timeliness: Up-to-date and relevant
Representativeness: Reflects target population
Fairness: Balanced across demographic groups

Bias Detection:

Underrepresentation of groups
Historical bias in labels
Proxy discrimination (e.g., zip code for race)
Sampling bias
Measurement bias

Privacy Compliance (GDPR/ISO 42001):

Lawful basis for processing (consent, legitimate interest, etc.)
Data subject rights supported (access, deletion, portability)
Privacy by design principles
Data Protection Impact Assessment (DPIA) if high-risk
Data Processing Agreements (DPAs) with vendors

STAGE 3: Model Development

Evaluate:

Appropriate algorithm selection
Explainability requirements considered
Fairness constraints incorporated
Robustness testing planned
Version control for code and models
Reproducibility ensured

Model Development Best Practices:

Baseline Establishment
- Simple model first (logistic regression, decision tree)
- Benchmark against human performance
- Justify complexity increase
Fairness Considerations
- Define fairness metrics (demographic parity, equalized odds, etc.)
- Test across protected attributes
- Trade-offs between accuracy and fairness documented
Explainability
- Use interpretable models when possible
- Apply XAI techniques (SHAP, LIME) for black-box models
- Document feature importance
- Provide example-based explanations
Adversarial Robustness
- Test against adversarial examples
- Implement input validation
- Monitor for distribution shift
Reproducibility
- Random seeds set
- Hyperparameters logged
- Environment documented (dependencies, versions)
- Training data snapshots preserved

STAGE 4: Validation and Testing

Evaluate:

Comprehensive test suite executed
Performance across subgroups validated
Fairness metrics measured
Robustness testing (adversarial, edge cases)
Safety and security testing
User acceptance testing (UAT)
Independent validation (if high-risk)

Testing Checklist:

Performance Testing:

Accuracy on test set
Precision, recall, F1-score
Performance by demographic group
Performance on edge cases
Calibration (confidence vs. accuracy)

Fairness Testing:

Demographic parity (equal acceptance rates)
Equalized odds (equal false positive/negative rates)
Predictive parity (equal precision)
Individual fairness (similar individuals treated similarly)

Robustness Testing:

Adversarial examples resistance
Input perturbation sensitivity
Out-of-distribution detection
Stress testing (high load, edge cases)

Safety Testing:

Failure mode analysis
Fallback mechanisms tested
Human override tested
Emergency stop procedures

Security Testing:

Model extraction attacks
Data poisoning resistance
Backdoor detection
Privacy leakage testing (membership inference)

Validation Outcome:

Pass: [Meets all criteria]
Conditional: [Meets most, some improvements needed]
Fail: [Major gaps, do not deploy]

STAGE 5: Deployment

Evaluate:

Phased rollout plan (pilot → limited → full)
Monitoring infrastructure in place
Human oversight mechanisms established
Incident response plan ready
User training and communication completed
Rollback plan prepared

Deployment Best Practices:

Pilot Testing
- Small user group
- Controlled environment
- Close monitoring
- Rapid feedback loops
Gradual Rollout
- Canary deployment (1% → 10% → 50% → 100%)
- A/B testing against baseline
- Monitor for unexpected impacts
Human-in-the-Loop
- Human review of high-stakes decisions
- Override capabilities
- Escalation procedures
- Audit sampling
Communication
- Notify affected users
- Provide transparency (AI disclosure)
- Explain rights and remedies
- Offer feedback channels

Deployment Checklist:

Infrastructure ready (compute, storage, APIs)
Monitoring dashboards configured
Alerting thresholds set
Incident response team trained
Legal and compliance approval obtained
Stakeholder communication sent
Documentation updated

STAGE 6: Monitoring and Maintenance

Evaluate:

Continuous performance monitoring
Drift detection (data and model)
Fairness monitoring over time
User feedback collection
Incident tracking and resolution
Regular model retraining
Audit trails maintained

Monitoring Framework:

1. Performance Monitoring

Accuracy, precision, recall (daily/weekly)
Latency and throughput
Error rates and types
Service availability (uptime)

2. Fairness Monitoring

Outcome disparities across groups (weekly/monthly)
False positive/negative rates by demographics
User satisfaction by group
Complaint rates

3. Data Drift Detection

Input distribution changes
Feature importance shifts
Anomaly detection
Trigger for retraining

4. Model Drift Detection

Prediction distribution changes
Confidence score patterns
A/B test against updated models

5. Safety Monitoring

Near-miss incidents
Human override frequency
Fallback activations
Edge case occurrences

Alert Triggers:

Accuracy drops > 5%
Fairness disparity exceeds threshold
Data drift detected
Error rate spike
Security anomalies
User complaints increase

Maintenance Schedule:

Daily: Dashboard review, alert triage
Weekly: Performance deep-dive, fairness check
Monthly: Model health assessment, incident review
Quarterly: Comprehensive audit, retraining evaluation
Annually: Full ISO 42001 compliance review

STAGE 7: Decommissioning

Evaluate:

Decommissioning criteria defined
Data retention/deletion policies
User migration plan (if replacement system)
Impact assessment of discontinuation
Archival and documentation
Lessons learned captured

Decommissioning Triggers:

End of useful life
Better alternative available
Regulatory prohibition
Unacceptable risk identified
Business need eliminated

Decommissioning Process:

Stakeholder notification (advance warning)
Gradual phase-out
Data handling (delete, anonymize, or archive)
Model archival (for audits)
Post-mortem analysis
Knowledge transfer

8.1 运行规划与控制

ISO 42001要求对AI全生命周期进行管理：

AI生命周期阶段：

设计 → 开发 → 验证 → 部署 → 监控 → 维护 → 退役

阶段1：设计与需求

评估：

清晰定义问题与成功标准
评估利益相关者需求
早期识别伦理考量
映射监管要求
开展可行性与影响分析
考虑AI替代方案

问题：

AI是否是合适的解决方案，或更简单的方法是否可行？
可能出现哪些问题？
谁会受到影响，如何影响？
需要哪些数据，是否可用？
存在哪些伦理红线？

警示信号：

在无正当理由的情况下，使用AI进行高风险决策
无明确的成功指标
忽视利益相关者担忧
数据不足或数据源存在偏见

阶段2：数据管理

评估：

评估数据质量（准确性、完整性、及时性）
分析偏见与代表性
记录与验证数据源
满足隐私与同意要求
数据安全与访问控制
应用数据最小化原则

数据质量维度：

准确性：正确且无错误
完整性：关键字段无缺失值
一致性：跨数据源统一
及时性：最新且相关
代表性：反映目标人群
公平性：不同人群间平衡

偏见检测：

群体代表性不足
标签存在历史偏见
代理歧视（如用邮政编码替代种族）
抽样偏见
测量偏见

隐私合规（GDPR/ISO 42001）：

具备合法的数据处理基础（同意、合法利益等）
支持数据主体权利（访问、删除、可携带性）
遵循隐私设计原则
高风险AI需开展数据保护影响评估（DPIA）
与供应商签订数据处理协议（DPA）

阶段3：模型开发

评估：

选择合适的算法
考虑可解释性要求
纳入公平性约束
规划鲁棒性测试
对代码与模型进行版本控制
确保可复现性

模型开发最佳实践：

基准建立
- 先构建简单模型（逻辑回归、决策树）
- 与人类绩效进行基准对比
- 说明增加复杂度的理由
公平性考量
- 定义公平性指标（人口均等、平等机会等）
- 针对受保护属性进行测试
- 记录准确性与公平性的权衡
可解释性
- 尽可能使用可解释模型
- 对黑盒模型应用XAI技术（SHAP、LIME）
- 记录特征重要性
- 提供基于示例的解释
对抗鲁棒性
- 测试对抗样本抗性
- 实施输入验证
- 监控分布偏移
可复现性
- 设置随机种子
- 记录超参数
- 文档化环境（依赖项、版本）
- 保留训练数据快照

阶段4：验证与测试

评估：

执行全面测试套件
验证跨子群体的性能
测量公平性指标
鲁棒性测试（对抗性、边缘案例）
安全与可靠性测试
用户接受度测试（UAT）
高风险AI需独立验证

测试清单：

性能测试：

测试集准确率
精确率、召回率、F1分数
不同人群的性能
边缘案例性能
校准度（置信度与准确率）

公平性测试：

人口均等（同等接受率）
平等机会（同等假阳性/阴性率）
预测均等（同等精确率）
个体公平性（相似个体得到相似对待）

鲁棒性测试：

对抗样本抗性
输入扰动敏感性
分布外检测
压力测试（高负载、边缘案例）

安全测试：

故障模式分析
fallback机制测试
人工覆盖测试
紧急停止流程

安全性测试：

模型提取攻击测试
数据投毒抗性
后门检测
隐私泄露测试（成员推理）

验证结果：

通过：[满足所有标准]
有条件通过：[满足大部分标准，需部分改进]
不通过：[存在重大差距，不得部署]

阶段5：部署

评估：

分阶段部署计划（试点 → 有限范围 → 全面部署）
监控基础设施到位
建立人类监督机制
事件响应计划就绪
完成用户培训与沟通
准备回滚计划

部署最佳实践：

试点测试
- 小范围用户群体
- 受控环境
- 密切监控
- 快速反馈循环
逐步部署
- 金丝雀部署（1% → 10% → 50% → 100%）
- 与基线进行A/B测试
- 监控意外影响
人类在环
- 人工审查高风险决策
- 覆盖能力
- 升级流程
- 审计抽样
沟通
- 通知受影响用户
- 提供透明度（AI披露）
- 解释权利与救济途径
- 提供反馈渠道

部署清单：

基础设施就绪（计算、存储、API）
监控仪表板配置完成
设置告警阈值
事件响应团队完成培训
获得法律与合规批准
发送利益相关者沟通信息
更新文档

阶段6：监控与维护

评估：

持续性能监控
漂移检测（数据与模型）
长期公平性监控
用户反馈收集
事件跟踪与解决
定期模型重新训练
维护审计轨迹

监控框架：

1. 性能监控

准确率、精确率、召回率（每日/每周）
延迟与吞吐量
错误率与类型
服务可用性（正常运行时间）

2. 公平性监控

不同群体的结果差异（每周/每月）
不同人群的假阳性/阴性率
不同群体的用户满意度
投诉率

3. 数据漂移检测

输入分布变化
特征重要性转移
异常检测
触发重新训练的条件

4. 模型漂移检测

预测分布变化
置信度得分模式
与更新模型进行A/B测试

5. 安全监控

未遂事件
人工覆盖频率
Fallback激活情况
边缘案例发生情况

告警触发条件：

准确率下降>5%
公平性差异超过阈值
检测到数据漂移
错误率激增
安全异常
用户投诉增加

维护计划：

每日：查看仪表板，处理告警
每周：深入分析性能，检查公平性
每月：模型健康评估，事件回顾
每季度：全面审计，评估重新训练需求
每年：完整的ISO 42001合规审查

阶段7：退役

评估：

定义退役标准
数据保留/删除政策
用户迁移计划（如有替代系统）
评估退役影响
归档与文档记录
总结经验教训

退役触发因素：

使用寿命结束
有更好的替代方案
被监管禁止
识别出不可接受的风险
业务需求消失

退役流程：

通知利益相关者（提前预警）
逐步淘汰
数据处理（删除、匿名化或归档）
模型归档（用于审计）
事后分析
知识转移

Step 6: Performance Evaluation (20 minutes)

步骤6：绩效评价（20分钟）

Clause 9: Performance Evaluation

条款9：绩效评价

Step 7: Improvement (15 minutes)

步骤7：改进（15分钟）

Clause 10: Improvement

条款10：改进

10.1 Nonconformity and Corrective Action

Evaluate:

Process for identifying nonconformities
Root cause analysis conducted
Corrective actions implemented
Effectiveness verified
AIMS updated to prevent recurrence

Example Nonconformities:

Fairness threshold breached
Undocumented model change
Training data bias discovered
Incident response delayed
Audit finding not addressed

Corrective Action Process:

Identify nonconformity
Immediate containment (stop harm)
Root cause analysis (5 Whys, Fishbone)
Corrective action plan
Implementation
Verification of effectiveness
Documentation and communication

10.2 Continual Improvement

Evaluate:

Process for ongoing improvement
Lessons learned captured
Best practices shared
Innovation encouraged
Benchmarking against industry

Improvement Opportunities:

New techniques for bias mitigation
Enhanced explainability methods
Automation of monitoring
Better stakeholder engagement
Process efficiency gains

Improvement Cycle:

Plan → Do → Check → Act (PDCA)

Apply continuously to AI systems and governance processes.

10.1 不符合项与纠正措施

评估：

具备识别不符合项的流程
开展根本原因分析
实施纠正措施
验证有效性
更新AIMS以防止复发

不符合项示例：

违反公平性阈值
模型变更未文档化
发现训练数据存在偏见
事件响应延迟
审计发现未处理

纠正措施流程：

识别不符合项
立即遏制（停止损害）
根本原因分析（5Why、鱼骨图）
纠正措施计划
实施
验证有效性
文档记录与沟通

10.2 持续改进

评估：

具备持续改进流程
收集经验教训
分享最佳实践
鼓励创新
与行业基准对比

改进机会：

偏见缓解新技术
增强可解释性方法
监控自动化
更好的利益相关者参与
流程效率提升

改进循环：

计划 → 执行 → 检查 → 处理（PDCA）

持续应用于AI系统与治理流程。

Complete ISO 42001 Audit Report

完整ISO 42001审计报告

markdown

undefined

markdown

undefined

ISO 42001 AI Governance Audit Report

ISO 42001 AI治理审计报告

AI System: [Name] Organization: [Name] Date: [Date] Auditor: [AI Agent] Standard: ISO/IEC 42001:2023

AI系统：[名称] 组织：[名称] 日期：[日期] 审计员：[AI Agent] 标准：ISO/IEC 42001:2023

Executive Summary

执行摘要

Compliance Status

合规状态

Overall Conformance: [Conformant / Partially Conformant / Non-Conformant]

Conformance by Clause:

Clause	Title	Status	Score	Critical Gaps
4	Context	✅ / ⚠️ / ❌	[X]/10	[List]
5	Leadership	✅ / ⚠️ / ❌	[X]/10	[List]
6	Planning	✅ / ⚠️ / ❌	[X]/10	[List]
7	Support	✅ / ⚠️ / ❌	[X]/10	[List]
8	Operation	✅ / ⚠️ / ❌	[X]/10	[List]
9	Evaluation	✅ / ⚠️ / ❌	[X]/10	[List]
10	Improvement	✅ / ⚠️ / ❌	[X]/10	[List]

Overall Score: [X]/100

总体合规性：[合规 / 部分合规 / 不合规]

按条款合规性：

条款	标题	状态	得分	重大差距
4	环境	✅ / ⚠️ / ❌	[X]/10	[列表]
5	领导力	✅ / ⚠️ / ❌	[X]/10	[列表]
6	规划	✅ / ⚠️ / ❌	[X]/10	[列表]
7	支持	✅ / ⚠️ / ❌	[X]/10	[列表]
8	运行	✅ / ⚠️ / ❌	[X]/10	[列表]
9	评价	✅ / ⚠️ / ❌	[X]/10	[列表]
10	改进	✅ / ⚠️ / ❌	[X]/10	[列表]

总体得分：[X]/100

Risk Classification

风险分类

AI System Risk Level: High / Limited / Minimal / Unacceptable

Justification: [Based on EU AI Act criteria and impact assessment]

AI系统风险等级：高 / 有限 / 低 / 不可接受

理由：[基于《欧盟AI法案》标准与影响评估]

Top 5 Critical Findings

五大关键发现

[Finding] - Clause [X] - Severity: Critical
- Risk: [Description]
- Impact: [Consequences]
- Recommendation: [Immediate action]
[Finding] - Clause [X] - Severity: High [Continue...]

[发现] - 条款[X] - 严重程度：重大
- 风险：[描述]
- 影响：[后果]
- 建议：[立即行动]
[发现] - 条款[X] - 严重程度：高 [继续...]

Positive Highlights

积极亮点

✅ [Strength 1]
✅ [Strength 2]
✅ [Strength 3]

✅ [优势1]
✅ [优势2]
✅ [优势3]

Detailed Findings

详细发现

[Full analysis by clause with evidence, gaps, and recommendations]

[按条款的全面分析，含证据、差距与建议]

Risk Assessment Summary

风险评估摘要

Critical Risks Identified

识别的重大风险

Risk 1: [Name]

Category: Ethical / Technical / Legal / Operational
Likelihood: High
Impact: Critical
Risk Level: CRITICAL
Current Controls: [Insufficient]
Required Actions: [List]
Owner: [Responsible party]
Deadline: [Date]

[Continue for all critical and high risks...]

风险1：[名称]

类别：伦理 / 技术 / 法律 / 运营
可能性：高
影响：重大
风险等级：重大
当前控制措施：[不足]
所需行动：[列表]
负责人：[责任方]
截止日期：[日期]

[继续列出所有重大与高风险...]

Compliance Roadmap

合规路线图

Phase 1: Critical Compliance (0-3 months)

阶段1：关键合规（0-3个月）

Objective: Address critical gaps and establish baseline compliance

Actions:

[Action 1] - Owner: [Name] - Due: [Date]
[Action 2] - Owner: [Name] - Due: [Date]
[Action 3] - Owner: [Name] - Due: [Date]

Success Criteria: [Measurable outcomes]

Investment: [Time, resources, budget]

目标：解决重大差距，建立基线合规性

行动：

[行动1] - 负责人：[名称] - 截止日期：[日期]
[行动2] - 负责人：[名称] - 截止日期：[日期]
[行动3] - 负责人：[名称] - 截止日期：[日期]

成功标准：[可衡量结果]

投入：[时间、资源、预算]

Phase 2: Enhanced Governance (3-6 months)

阶段2：强化治理（3-6个月）

Objective: Strengthen AI governance and risk management

Actions: [List...]

目标：加强AI治理与风险管理

行动： [列表...]

Phase 3: Maturity and Optimization (6-12 months)

阶段3：成熟与优化（6-12个月）

Objective: Achieve full conformance and continual improvement

Actions: [List...]

目标：实现全面合规与持续改进

行动： [列表...]

Documentation Requirements

文档要求

Missing Documentation

缺失的文档

Recommendations by Stakeholder

按利益相关者划分的建议

For Leadership

针对领导层

Establish AI Ethics Committee
Allocate budget for responsible AI
Mandate ISO 42001 compliance

建立AI伦理委员会
为负责任AI分配预算
强制要求ISO 42001合规

For AI Teams

针对AI团队

Implement fairness testing in CI/CD
Create model cards for all systems
Conduct bias audits quarterly

在CI/CD中实施公平性测试
为所有系统创建模型卡片
每季度开展偏见审计

For Legal/Compliance

针对法律/合规团队

Monitor regulatory developments (EU AI Act)
Update privacy policies for AI use
Establish DPIA process for high-risk AI

监控监管动态（《欧盟AI法案》）
更新AI使用相关隐私政策
为高风险AI建立DPIA流程

For Operations

针对运营团队

Deploy monitoring infrastructure
Implement human oversight mechanisms
Create incident response runbooks

部署监控基础设施
实施人类监督机制
创建事件响应手册

Next Steps

下一步行动

Appendices

附录

A. ISO 42001 Checklist

A. ISO 42001清单

[Detailed requirement-by-requirement checklist]

[按要求逐条列出的详细清单]

B. Risk Register

B. 风险登记册

[Complete risk inventory with assessments]

[完整风险清单及评估]

C. Glossary

C. 术语表

[AI and ISO terminology]

[AI与ISO术语]

D. References

D. 参考文献

ISO/IEC 42001:2023
EU AI Act
NIST AI Risk Management Framework
[Industry-specific standards]

Report Version: 1.0 Confidentiality: [Internal / Confidential / Public]

---

ISO/IEC 42001:2023
《欧盟AI法案》
NIST AI风险管理框架
[行业特定标准]

报告版本：1.0 保密性：[内部 / 机密 / 公开]

---

ISO 42001 Compliance Checklist

ISO 42001合规清单

Use this quick reference for self-assessment:

使用本快速参考进行自我评估：

Clause 4: Context ✓

条款4：环境 ✓

AIMS scope defined
Stakeholders identified
External issues (regulatory, social) assessed
Internal capabilities evaluated

定义AIMS范围
识别利益相关者
评估外部问题（监管、社会）
评估内部能力

Clause 5: Leadership ✓

条款5：领导力 ✓

Management commitment documented
AI policy established
Roles and responsibilities assigned
AI ethics committee or similar

记录管理层承诺
建立AI政策
分配角色与职责
成立AI伦理委员会或类似机构

Clause 6: Planning ✓

条款6：规划 ✓

AI objectives set
Risk assessment conducted
Risk treatment plans documented
Opportunities for improvement identified

设定AI目标
开展风险评估
记录风险处理计划
识别改进机会

Clause 7: Support ✓

条款7：支持 ✓

Resources allocated (compute, budget, people)
Competence requirements defined
Training provided
Awareness program active
Documentation maintained

Clause 8: Operation ✓

条款8：运行 ✓

Clause 9: Evaluation ✓

条款9：评价 ✓

Performance monitoring
Internal audits scheduled
Management reviews conducted
KPIs tracked

绩效监控
安排内部审计
开展管理层审查
跟踪KPI

Clause 10: Improvement ✓

条款10：改进 ✓

Nonconformity process
Corrective actions
Continual improvement culture

不符合项流程
纠正措施
持续改进文化

Best Practices

最佳实践

Start with Risk Assessment: Prioritize based on AI risk level
Document Everything: ISO 42001 requires extensive documentation
Engage Stakeholders Early: Include affected parties in governance
Use Existing Frameworks: Leverage NIST AI RMF, EU AI Act requirements
Automate Monitoring: Build MLOps with governance built-in
Train Your Team: ISO 42001 requires competent personnel
Regular Audits: Don't wait for problems—proactive reviews
Learn from Incidents: Every issue is improvement opportunity
Balance Innovation and Safety: Responsible AI doesn't mean no AI
Seek Certification: Third-party ISO 42001 certification adds credibility

从风险评估开始：根据AI风险等级确定优先级
记录所有内容：ISO 42001要求大量文档
尽早参与利益相关者：将受影响方纳入治理
利用现有框架：借鉴NIST AI RMF、《欧盟AI法案》要求
自动化监控：构建内置治理的MLOps
培训团队：ISO 42001要求具备能力的人员
定期审计：不要等问题出现——主动审查
从事件中学习：每个问题都是改进机会
平衡创新与安全：负责任AI不意味着拒绝AI
寻求认证：第三方ISO 42001认证提升可信度

Regulatory Alignment

监管对齐

ISO 42001 aligns with major AI regulations:

EU AI Act:

Risk classification framework
High-risk AI obligations
Transparency requirements
Conformity assessment

GDPR:

Data protection by design
Privacy impact assessments
Data subject rights
Lawful processing

NIST AI RMF:

Govern, Map, Measure, Manage functions
Risk-based approach
Trustworthy AI characteristics

Sector-Specific:

Healthcare: FDA AI/ML guidance, MDR
Finance: Model Risk Management (SR 11-7)
Employment: EEOC AI guidance

ISO 42001与主要AI法规对齐：

《欧盟AI法案》：

风险分类框架
高风险AI义务
透明度要求
合规评估

GDPR：

隐私设计
隐私影响评估
数据主体权利
合法处理

NIST AI RMF：

治理、映射、测量、管理功能
基于风险的方法
可信AI特征

行业特定：

医疗：FDA AI/ML指南、MDR
金融：模型风险管理（SR 11-7）
就业：EEOC AI指南

Common Pitfalls

常见陷阱

"We'll add governance later" - Build it in from the start
Treating ISO 42001 as one-time exercise - It's continual
Documentation without implementation - Must be operational
Ignoring low-risk AI - Even minimal-risk needs baseline governance
No stakeholder engagement - Affected parties must be involved
Insufficient resources - Responsible AI requires investment
Lack of monitoring - Deploy-and-forget is non-compliant
No incident response plan - When AI fails, you need a plan
Training as checkbox - Teams must truly understand responsible AI
Copying templates without customization - Tailor to your context

“以后再添加治理” - 从一开始就内置治理
将ISO 42001视为一次性工作 - 这是持续的过程
有文档但未执行 - 必须可操作
忽视低风险AI - 即使低风险也需要基线治理
无利益相关者参与 - 必须纳入受影响方
资源不足 - 负责任AI需要投入
监控不足 - 部署后就不管不符合合规要求
无事件响应计划 - AI故障时需要应对方案
培训走过场 - 团队必须真正理解负责任AI
照搬模板而不定制 - 需根据自身情况调整

Version

版本

1.0 - Initial release based on ISO/IEC 42001:2023

Remember: ISO 42001 is about building trustworthy AI systems through systematic risk management and governance. It's not a barrier to innovation—it's a framework for responsible innovation that protects both organizations and the people affected by AI.

1.0 - 基于ISO/IEC 42001:2023的初始版本

请记住：ISO 42001旨在通过系统的风险管理与治理，构建可信AI系统。它不是创新的障碍——而是负责任创新的框架，同时保护组织与受AI影响的人群。

iso-42001-ai-governance

Original

Translation

ISO 42001 AI Governance Audit

ISO 42001 AI治理审计

When to Use This Skill

何时使用本技能

Inputs Required

所需输入

ISO 42001 Framework Overview

ISO 42001框架概述

Core Clauses

核心条款

Key ISO 42001 Principles

ISO 42001核心原则

1. Risk-Based Approach

1. 基于风险的方法

2. Ethical AI

2. 伦理AI

3. Lifecycle Management

3. 全生命周期管理

4. Stakeholder Engagement

4. 利益相关者参与

Audit Procedure

审计流程

Step 1: Context and Scope Analysis (15 minutes)

步骤1：环境与范围分析（15分钟）

Step 2: Leadership and Governance Evaluation (20 minutes)

步骤2：领导力与治理评估（20分钟）

Clause 5: Leadership

条款5：领导力

Step 3: Planning and Risk Management (30 minutes)

步骤3：规划与风险管理（30分钟）

Clause 6: Planning

条款6：规划

Risk: [Name]

风险：[名称]

Step 4: Support and Resources (20 minutes)

步骤4：支持与资源（20分钟）

Clause 7: Support

条款7：支持

Step 5: Operation - AI Lifecycle Management (40 minutes)

步骤5：运行 - AI全生命周期管理（40分钟）

Clause 8: Operation

条款8：运行

Step 6: Performance Evaluation (20 minutes)

步骤6：绩效评价（20分钟）

Clause 9: Performance Evaluation

条款9：绩效评价

Step 7: Improvement (15 minutes)

步骤7：改进（15分钟）

Clause 10: Improvement

条款10：改进

Complete ISO 42001 Audit Report

完整ISO 42001审计报告

ISO 42001 AI Governance Audit Report

ISO 42001 AI治理审计报告

Executive Summary

执行摘要

Compliance Status

合规状态

Risk Classification

风险分类

Top 5 Critical Findings

五大关键发现

Positive Highlights

积极亮点

Detailed Findings

详细发现

Risk Assessment Summary

风险评估摘要

Critical Risks Identified

识别的重大风险

Compliance Roadmap

合规路线图

Phase 1: Critical Compliance (0-3 months)

阶段1：关键合规（0-3个月）

Phase 2: Enhanced Governance (3-6 months)

阶段2：强化治理（3-6个月）

Phase 3: Maturity and Optimization (6-12 months)